dquail / predictiveknowledge Goto Github PK

def ta_learn(lastState, action, newState):
            return

def distanceToTouchAdjacentCumulant(phi):
            return 1.0

def gtdLearn(self, lastState, action, newState):
        pred = self.prediction(lastState)
        zNext = self.cumulant(newState)
        gammaNext = self.gamma(newState)
        lam = self.lam(newState)
        rho = self.rho(action, lastState)
        self.eligibilityTrace = rho * (self.gammaLast * lam * self.eligibilityTrace + lastState)
        newStateValue = 0.0
        if not newState  is None:
            newStateValue = np.inner(newState, self.weights)
        tdError = zNext + gammaNext * newStateValue - np.inner(lastState, self.weights)
        updateH = self.alphaH * (tdError * self.eligibilityTrace - (np.inner(self.hWeights, lastState)) * lastState)
        self.hWeights = self.hWeights + updateH
        self.i = self.i + 1

        upWeights = self.alpha * (tdError * self.eligibilityTrace - gammaNext * (1-lam)  * (np.inner(self.eligibilityTrace, self.hWeights) * newState))
        if (zNext >0):
            t = 0
            for w in upWeights:
                if w>0:
                    t = t + 1
        self.weights = self.weights + upWeights
        pred = self.prediction(lastState)
        self.gammaLast = gammaNext

self.ACTIONS = {
            'forward': "forward",
            'turn_left': "turn_left",
            'turn_right': "turn_right",
            'extend_hand': "extend_hand"
        }